How agentic AI is rewiring Amazon’s teams and upending its traditions

[Editor’s Note: Agents of Transformation is an independent GeekWire series, underwritten by Accenture, exploring the adoption and impact of AI and agents. See coverage of our related event.]

Amazon is legendary for its process of “working backwards.” Start with a customer problem, imagine a future in which it’s solved, draft a press release and FAQs as if it had already happened, obsess over the document until it’s just right, and then go make it a reality.

But sometime last year, it dawned on Swami Sivasubramanian, Amazon Web Services VP of agentic AI, that new coding tools had suddenly made it easier for his teams to develop a demo — actual working software — than to write and perfect the classic six-page Amazon “PRFAQ.”

So they began starting with the prototype instead.

If something is “a low-risk bet where we just want to prove our intuition, then I actually say, let’s first go build the demo, and then iterate,” Sivasubramanian said in an interview last week, in advance of his keynote address Wednesday at the AWS New York Summit.

It’s an illustration of how agentic tools are reshaping even the most entrenched workplace practices and traditions. But it’s just one of the ways that the AWS agentic AI team is departing from the company’s established norms, and in some ways returning to its roots. 

Inside Amazon, CEO Andy Jassy says he wants the company to run like the world’s largest startup. Sivasubramanian’s division may be the closest thing to what that looks like in practice. 

Back to two pizzas

The AWS agentic AI division is organized into dozens of small teams, many of them just large enough to feed with two pizzas. That was the organizing principle that Amazon pioneered in its early days and that much of the company outgrew as it scaled to 1.5 million employees. 

When Matt Garman, the CEO of AWS, carved out agentic AI as its own division last year, Sivasubramanian went with small teams on purpose. It matches the new reality of the AI era: projects that once required 30 to 40 people, he said, can now be done by teams of six to eight.

Case in point: the Amazon Quick desktop app, which connects to a user’s email, calendar, Slack, documents, and other apps in a single workspace, and uses AI to search across them, answer questions, and perform tasks. It’s Amazon’s entry in a market where Anthropic, Microsoft, Google, and OpenAI have captured much of the attention.

It traces its roots to late January of this year, when Sivasubramanian said it became clear to him and others on the team that the underlying models had gotten good enough that the main missing ingredient was connecting them to the systems where people actually work. 

He pulled together a team of about six engineers to build it. Six weeks later, 200 people inside Amazon were using it. Ten weeks in, it was up to 10,000 internally. The team circled back to write the PRFAQ after the product was already in beta, to help refine their approach to the external launch. They shipped on April 28, three months after they got started.

Under the old system — writing the PRFAQ, routing it through layers of review — the paperwork alone could have taken as long as building and shipping the actual product.

Similar stories are playing out across the division. 

• One team open-sourced Strands, an AWS software development kit for building AI agents, after a member of Sivasubramanian’s team messaged him at 7 a.m. with the idea. After a quick call with Garman, they decided to go ahead. Within days, it was done.
• Kiro, the AI coding tool, was built by a deliberately small team, using Kiro itself to build it. One engineer prototyped a complex cross-platform notification feature for Kiro that had been estimated at four weeks of work, and shipped it in a day and a half. 
• The internal Amazon team that rebuilt the inference engine for the company’s Bedrock platform for AI models did it with six engineers in 76 days, a project originally expected to take 30 developers 12 to 18 months. 

Smaller teams everywhere

What’s happening inside Amazon’s agentic AI division is part of a trend across the tech industry toward smaller teams and flatter organizations, driven by AI and agents. 

Microsoft’s 2026 Work Trend Index, a survey of 20,000 workers in 10 countries, found that the biggest factor behind AI’s real impact in the workplace isn’t individual skill but whether the organization has restructured around the new technologies. 

Vijaye Raji, OpenAI’s CTO of applications, said during a recent Technology Alliance event that the company’s “ambitions are growing faster than we can hire people” — but the profile of who gets hired is changing. OpenAI increasingly looks for engineers who work with AI tools natively, and the gap between those who do and those who don’t is stark: the top engineers at OpenAI use roughly 100 times more AI tokens than the median.

All of this leads to a natural question: what does this mean for jobs? Amazon has cut roughly 30,000 corporate jobs since late 2025 as part of what Jassy has described as an effort to reduce bureaucracy. He has said he expects AI to shrink the corporate workforce over time. 

Similar cuts are playing out across the industry, from Meta to Block to LinkedIn, as companies rethink not only the roles they need to fill but also how many people they need overall. 

Bigger goals, same team

Sivasubramanian describes the shift differently: In his division, the same number of people are now pursuing a bigger charter. With the new structure, they’re able to take on more projects, and faster, accomplishing things in weeks that would have taken much longer in the past.

The nature of the roles inside those teams is changing, too. Increasingly, product managers write code, and engineers make product decisions. On the Kiro team, for example, a product manager built the first version of a cost analysis dashboard using Kiro itself. 

This also requires leaders to operate differently. For example, Sivasubramanian said he is careful to monitor which decisions need his approval, even when traveling. At the current pace, even four or five days of delay can add as much as 10% to a team’s shipping timeline. 

Managing these teams also raises new questions. Sivasubramanian said his division has started tracking how much it spends on AI tokens — the basic unit of interaction with an AI model — the way it would track any other operating cost. 

So far, the numbers have been manageable: tools like Kiro invest upfront in defining specs and pulling in the right context before generating code, which makes them more efficient with tokens rather than burning through them in aimless back-and-forth. 

Even the heaviest users consume only a few thousand dollars a month, he said. But he expects that over time, companies will need a full picture of their operating expenses that includes not just headcount but the cost of the AI agents working alongside them.

This gets to a bigger point: “The bottleneck is not about the time it takes to build something,” Sivasubramanian said. “The bottleneck is about crafting the right specification and the tests and the right product and customer experience.”

In a blog post published last week, Sivasubramanian wrote that teams across the company that restructured their workflows around AI saw a median 4.5x productivity gain, with some exceeding 10x gains. The teams that simply added AI tools to their existing way of working didn’t see the same results.

Coding and testing

That shift has created its own challenges. Teams can generate code faster than ever, but if they don’t define what success looks like up front — the specs, the tests, the edge cases — the agents don’t have as much chance of success. 

Amazon is now pushing testing to the moment of coding rather than handling it in stages, so agents can check their own work before anything reaches production.

Sivasubramanian learned this first-hand, the hard way. Earlier this year, jet-lagged and unable to sleep in his hotel room on a trip to India, he decided to try a fun project: He used Kiro to rebuild a piece of AWS infrastructure he’d originally developed by hand nearly 20 years ago — a replication engine that still underpins core services like S3 and DynamoDB.

He and one of Amazon’s earliest distinguished engineers, Allan Vermeulen, had spent four months on the original. Sivasubramanian figured the agent would make quick work of it. Instead, he spent four nights going back and forth, babysitting each step. 

On the fifth night, he realized the problem: he hadn’t given the agent the tools to test its own output. Once he wrote the right spec and set up the testing environment, it was done in about two hours. Asked what he did with his rebuilt version of the engine, Sivasubramanian laughed. He never shipped it. “Maybe I should have,” he said.

With the right team and a couple of pizzas, maybe he still can.